1

ANGELAH: A Framework for Assisting Elders AtHome

Tarik Taleb1,∗, Dario Bottazzi2,†, Mohsen Guizani3,‡, and Hammadi Nait-Charif4,§1Graduate School of Information Sciences, Tohoku University, Japan

2Department of Electronics, Computer Science, and System Engineering, University of Bologna, Italy3Department of Computer Science, Western Michigan University, USA

4National Centre for Computer Animation, Bournemouth University, UK∗ talebtarik@ieee.org, † dario.bottazzi@unibo.it, ‡ mguizani@ieee.org, § hncharif@bournemouth.ac.uk

Abstract—The ever growing percentage of elderly peoplewithin modern societies poses welfare systems under relevantstress. In fact, partial and progressive loss of motor, sensorial,and/or cognitive skills renders elders unable to live autonomously,eventually leading to their hospitalization. This results in bothrelevant emotional and economic costs.

Ubiquitous computing technologies can offer interesting op-portunities for in-house safety and autonomy. However, existingsystems partially address in-house safety requirements and typ-ically focus on only elder monitoring and emergency detection.The paper presents ANGELAH, a middleware-level solutionintegrating both ”elder monitoring and emergency detection”solutions and networking solutions. ANGELAH has two mainfeatures: i) it enables efficient integration between a variety ofsensors and actuators deployed at home for emergency detectionand ii) provides a solid framework for creating and managingrescue teams composed of individuals willing to promptly assistelders in case of emergency situations. A prototype of ANGE-LAH, designed for a case study for helping elders with visionimpairments, is developed and interesting results are obtainedfrom both computer simulations and a real-network testbed.

Index Terms—Ubiquitous assistance, pervasive computing,healthcare network, and elderly assistance.

I. INTRODUCTION

Major developed-world societies are experiencing a fright-ening demographic phenomenon of a rapid growth of theiraging population [1]. As the elderly people percentage ismounting, health and social costs are increasing too, thusputting welfare systems under relevant stress. In fact, pro-gressive degradation of vision, hearing, motion, and cognitiveskills deprives people from the possibility of independentlyperforming basic activities, such as, self-care, leisure, house-hold, and social interactions, and often forces elders’ hospital-ization, with both emotional and economic notable impacts.

Technological advances and cost reduction in computingdevices and network solutions can play an important role inenhancing elderly people independence. In particular, recentdevelopments in wireless technologies, sensors, and actuatorsare enabling new classes of eldercare applications availableanywhere and at anytime, i.e., ubiquitous eldercare services.The common guideline behind ubiquitous eldercare is thecomplete shift of the locus of health control from hospitalsto pervasive systems deployed close to where elderly userslive and move, with the main goals of increased independence,

safety, and quality of life on one hand, and of care cost-savingon the other hand. Along this line, several ubiquitous eldercaresolutions have recently appeared in the literature that permitto record and analyze elders behavioral patterns, to monitorseniors’ mobility, to assist individuals in daily activities, suchas self-care, and so forth.

However, to the best of our knowledge, only few solutionsaim at addressing domestic safety for elders [2], [3]. Individ-uals safety is a prerequisite for autonomous life-styles. Theincrease, each year in deaths and injuries, especially amongelders, has shown up “in-house safety” as an emergent fieldof research. In fact, everyday tasks are a continuous source ofdanger for elders due to their eventual decline in physical andcognitive skills.

Supporting in-house safety is a rather challenging task thatrequires suitable answers to several research questions:• how can we monitor and detect possibly dangerous situ-

ations?• how can we design unobtrusive assistive solutions that

facilitate integration, management and update of sensorsand actuators infrastructure within elders home environ-ments?

• how can we improve emergency detection and response?• and finally, how can we identify suitable trade-offs that

keep into account all of these considerations?The main motivation behind this research work emerges

from the conviction that the graceful integration of Com-mercial Off-The-Shelf (COTS) devices and networking el-ements, along with pervasive computing technologies, canoffer significant opportunities to provide safety and quality oflife to elderly in need of help. This research work presentsthe ANGELAH (AssistiNG ELders At Home) framework,a middleware-level solution able of i) integrating sensorsand actuators needed to monitor and guarantee elder safety,ii) detecting possibly dangerous situations for the elder, andiii) composing emergency response groups of volunteers andcaregivers, allocated in the nearby, willing to help in caseof an emergency event [3]. As will be explained later, theemergency response group members are chosen according toseveral criteria, such as their physical proximity to elders homeand their medical skills, following the multi-attribute decisionmaking (MADM) theory [4].

The remainder of this paper is organized as follows. Section

2

II sets the present work within the state of the art. SectionIII identifies the major requirements and design guidelinesthat should be followed to support in-house elder safetyservices. Section IV presents the ANGELAH model. SectionV portrays the system architecture. Section VI shows relevantimplementation insights on response group member selection.Section VII presents a case study of the proposed ANGELAHframework. Concluding remarks follow in Section VIII.

II. RELATED RESEARCH WORK

To better position the past and on-going research activitiesabout ubiquitous eldercare, let us first introduce a solutionclassification widely accepted in the medical area. The US Na-tional Center on Medical Rehabilitation Research (NCMRR)has proposed a model for categorizing the research and devel-opment efforts in technology for aging in well conditions [5].The NCMRR model identifies five main working directions atdifferent levels, each focusing on a specific problem: cellular,organ, action, task-role, and social limitations. The cellularlevel relates to aberrations in normal physiological processesand in the cellular structure, with the consequent diseasesand/or genetic abnormalities. The organ level focuses onsolutions to impairments of organs, e.g., the heart, or of wholeorgan systems, e.g., the cardio-vascular system, while theaction level refers to person inabilities to properly performsome actions due to functional limitations of the responsibleorgans. The research at the task-role level is addressed topeople disabilities in properly performing tasks and activitiesin specific physical and social contexts. The highest researcharea in the NCMRR model is about social barriers that obstaclepeople engagement in a rich socio-emotional environment.

Each NCMRR level involves specific technological require-ments and calls for proper solutions to enable elder autonomyin daily living, with less dependencies on family, friends, andcaregivers support. Let us note that only in the cellular areathe research focus is more on cure and recovery of agingpeople rather than on technological solutions to improve theirquality of life. More in details, research efforts in the organ andaction levels seek to design and develop devices to compensatelimitations in common physical functionality, such as move-ment or hearing. For example, advanced signal processingtechnologies have permitted to significantly improve the soundquality provided via digital hearing aids. Technologies atthe NCMRR task-role level aim at reducing the impact ofdisabilities into the aging people life. For instance, a hearingaid should also allow aging people with hearing disabilities toactively participate in usual conversations. Research efforts inthe social limitation level, instead, analyze the social attitudesof aging people to derive the guidelines for the development oftechnological solutions that can improve the social interactionsof elders with the external world.

Pervasive computing offers relevant and challenging oppor-tunities to design and implement ubiquitous eldercare solutionsat the task-role and social limitation levels [23]. In fact, on onehand, ubiquitous assistance solutions enable the realization ofwearable devices with ubiquitous connectivity that can operateat the task-role level by assisting elders in their activities and

by addressing their disabilities anywhere and anytime. On theother hand, ubiquitous care networking supports permit toovercome social limitations by providing elders with a plethoraof communication artifacts and services specifically tailoredto set them within the context of a rich social and emotionalframework and to reduce their sense of loneliness.

Ubiquitous assistance solutions. In the first category, therehas been a plethora of research work addressing various as-pects of telemedicine [6]. The scope of these researches rangefrom standardization activities for telemedicine deployment[7], privacy and security solutions [8] to realization of a system[9], [10], [11]. The main concerns of these solutions consistin assisting elders in their routine life activities, constantmonitoring of their health conditions, and prompt alerting inemergency events. In [12], a computer-vision based system isproposed to support people with severe vision impairments.To secure a safe navigation of a particular environment, thesystem generates alert messages, via a speech output interface,whenever a change occurs to the layout of the environment.This makes people, with vision deficiency and living in thatenvironment, aware of the occurring changes. In [13], artificialintelligence learning and planning techniques are used todefine proper steps of basic activities of daily living (e.g.,hand-washing). The resultant system provides visual or verbalinstructions to a person with dementia on how to perform aparticular daily living activity. The system consists of threemodules; a tracking module which uses computer vision tomonitor the actions of the user by determining the spatial coor-dinates of the person’s body and hands within the environment.Once these coordinates have been determined, a planning mod-ule determines what step the user is completing and whetherthe step being completed is correct. If the system detects thatthe user has made an error, such as completing a step out ofsequence or missing a step altogether, the prompting moduleselects and plays a prompt message. In [14], a cognitiveorthotic system, called auto-minder, is proposed. The systemmodels the daily plans of an individual and decides on whenand where to remind the person of the execution of those plans.The developed techniques are deployed on a mobile robot, aspart of the Nursebot project’s Initiative on Personal RoboticAssistance of the Elderly. Benny et al., in their UbiSense sys-tem, used embedded smart vision techniques to detect changesin posture, gait and activities. In addition to monitoring normaldaily activities and detecting potential adverse events suchas falls, the system aims to capture signs of deterioration ofthe patients by analyzing subtle changes in posture and gait[15]. In their later work, they used image sensing and vision-based reasoning to verify and further analyze events reportedby other sensors such as Accelerometer for fall detection[16]. The Honeywell Laboratories’ Independent Life StyleAssistant (ILSA) is another notable example of integratedsmart environments which aim at ubiquitous assistance [17].In the ILSA system, multiple JADE agents are deployed. Theagents support data monitoring via home-installed sensors. Thecollected data are aggregated and processed to make adequateresponse planning and machine learning. UbiMedic [18] andCode Blue [19] are other notable agent-based smart homeassistance systems. A general observation about the different

3

industrial and academic proposals of ubiquitous assistance isthat they focus on only specific tasks, and thus provide only asubset of the support functions needed for eldercare assistance.

Ubiquitous care networking. The ubiquitous care network-ing area is still in its infancy, with only a few solutionproposals. The primary objective of ubiquitous care network-ing solutions is to promote social interactions of elders withtheir surroundings and to coordinate between the differentactors involved in a ubiquitous eldercare event. In [20], acontext-aware middleware solution, dubbed Allocation andGroup Aware Pervasive Environment (AGAPE), is proposedfor the creation and management of ad-hoc assistance teamsto provide outdoor emergency assistance to senior citizens inneed of immediate help.

To conclude, different solutions have been devised in therecent literature for both ubiquitous assistance and ubiquitouscare networking support. To the best knowledge of the authors,no solution in the literature has integrated the two types ofubiquitous eldercare into one single solution. This challengingtask underpins the focus of the research work outlined in thispaper.

III. UBIQUITOUS IN-HOUSE SAFETY REQUIREMENTS ANDDESIGN GUIDELINES

The design of technological solutions for ubiquitous in-house safety is a rather challenging problem. Most of availableworks in the literature provide ad-hoc solutions for elder mon-itoring, provide tools to detect specific emergency situationsfor the elder, and typically provide basic support for alertinghealth authorities.

Elder monitoring is typically performed by exploiting sen-sors, such as, RFIDs, accelerometers, pressure sensors, video-cameras that are pervasively deployed in the environmentwhere the elder lives and/or embedded in the devices he1 cur-rently uses. However, most available research work typicallyfocuses on specific elder pathologies and can only partiallyaddress the requirements stemming from in-house safety. Sev-eral technological challenges have to be addressed to supportelder monitoring. In fact, notwithstanding the widespreadavailability of low-cost sensors of different types, sensorscurrently exhibit a high heterogeneity in nature, accuracy, andperformance and should be chosen according to elders’ needsand pathology. They should be unobtrusive and must not alterthe existing living environment; their monitored data should beas much interoperable as possible, so that different assistancetasks can exploit the context information deriving from thesame sensors; they might be changed during the time accord-ing to the progressive changes in elders’ needs [21], [22],[23]. In addition, sensor technologies for eldercare should notplace any additional load on final users by requiring explicitinput/effort from older adults [24]. Indeed, even rememberingto wear a device/sensor may represent a cognitive load that afraction of elder population cannot tolerate, e.g., elderly peopleaffected by dementia.

In case of emergency situations, available in-house safetysupports typically permit to alert only heath authorities, and

1In this paper, the term “he” also refers to she.

rarely friends or relatives. However, neighbors and passing-bypeople may play an important role in emergency situations.Indeed, prompt assistance may become the determining factorbetween life and death in case of heart attacks: circulationand ventilation should be initiated as soon as possible toavoid irreversible cerebral damage. However, high levels ofstress and lack of time, typically characterizing emergencysituations, could create untrained and unrehearsed reactionsthat cause critical delays in intervention and contribute toinconsistencies/omissions in the critical data to share. Novelcommunication/coordination solutions should proactively pro-pose and stimulate collaboration among potential helpers co-located with the elder in need of help: they should clearlynotify helper community members about the occurring situa-tion, should inform helpers about the activities to perform bysupporting them in the collaborative definition and assignmentof assistance tasks, and should coordinate their activities viasynchronization and data sharing, dynamically tailored to thegroup of participants and to the degree of criticality of theelder situation [3].

According to the above considerations, we need to considertwo main design principles when designing an in-house safetysupport solution: context-awareness and group-based collabo-ration.

Context-Awareness. The full visibility of the elder contextinformation [25], such as elder physical location, gestures,health status, and clinical record, permits to dynamically deter-mine whether the elder is in need of help. If so, to tailor serviceprovisioning according to the current operating conditions. Forinstance, in the case of an accidental fall the system can detectthe occurred event and may accordingly react by contactingelder’s relatives and caregivers, by triggering actuators tounlock the elder’s home doors. In addition, the visibility ofcontext information may also permit to tailor the interactionbetween the elder and the system, e.g., by adopting speech-based user interfaces in the case of individuals with strongvisual impairments.

Context information should be obtained by sensor infras-tructures composed on the basis of elder needs and patholo-gies. As a consequence, context information tend to show ahigh degree of heterogeneity, making it necessary to carefullyconsider mechanisms and tools to turn raw monitoring datafrom sensors into context information at a higher level of ab-straction, ready to be used by the application level. In particu-lar, to improve software re-usability and to accelerate in-housesafety service customization, it is necessary to decouple thedesign and implementation of context aggregation solutionsfrom the development and deployment of assistance services.This permits to hide assistance services from the technologicaldetails and difficulties of sensing and monitoring. For example,an emergency response service generally requires knowingwhen an elder is calling for prompt help, e.g., when he hasfallen, rather than the sampled value from a position sensor.In addition, to recognize that a complex context situation hasoccurred, it is often necessary to compare sensed values frommultiple sensors. Aggregation is the support functionality thatallows interpreting and reasoning on complex context situa-tions by putting together and processing the raw monitoring

4

data coming from sensors, generally increasing the abstractionlevel of context data.

Many-to-many context data distribution is another importantrequirement to take into account in the design of ubiqui-tous assistance services. On one hand, assistance applicationsneed to exploit context that may come from different, andoften heterogeneous, sources. On the other hand, the samecontext information should be provided to several distributedservice components, thus enabling different ubiquitous elder-care services to obtain the needed context information fromthe same sensor infrastructure. For instance, an integratedpositioning system may consist of indoor (Wi-Fi-based) andoutdoor (GPS-based) location detectors, possibly producingpositioning data in different formats. The result of their workshould be distributed to both an emergency service and anelder-tracking service for family members.

Group-based Collaboration. Grouping is the primary issueto address in the design of support mechanisms and tools forin-house safety. The metaphor of group, team, or communitycan facilitate interactions among individuals. In fact, the groupabstraction permits to restrict the scope of in-house safetyinformation dissemination/sharing, such as emergency alerts,to only the members of a group. This can significantly simplifythe establishment of inter-member interactions and decisionmaking. Various types of eldercare communities can be formedspontaneously to provide impromptu assistance without anyprior planning. In fact, group membership management for in-house safety scenarios cannot rely on preconceived knowledgeabout the set of available group members, their characteristics,and their device properties. On the contrary, it is necessary toprovide users with the possibility to compose, dissolve, join,and leave groups on-demand from highly heterogeneous termi-nals, with possibly different forms of wireless connectivity. Forinstance, teams consisting of individuals in physical proximitywith an elder in faint can be dynamically created to provideprompt assistance to the elder. These groups may includepeople with limited medical skills, professional caregivers,as well as doctors who are currently located in physicalproximity.

Context awareness is emerging as a crucial property inthe design of group management solutions for in-housesafety. Novel group supports should exploit the full visibilityof context information, such as elder conditions, locationof users/terminals/resources, user profile of preferences andskills, to promote the collaboration among members willingto assist the elder in-house anytime. Among the various andheterogeneous data belonging to context, location plays aprimary role. For instance, group management supports couldexploit the visibility of elder location to create an eldercaresupport group on the fly, composed of only the entities in theelder vicinity to provide prompt help. This may be particularlyrelevant in the case of emergency situations, such as heartattacks, where response operations are subject to severe timeconstraints. The visibility of further context information, suchas individuals’ medical skills, may also guide the formationof suitable ad-hoc groups: for example, medical skills maypermit to preferably compose response groups of people withappropriate experience to cope with the emergency situation.

In the proposed ANGELAH framework, different contextfactors are envisioned and MADM theory is applied to themto form adequate groups of volunteers.

IV. THE ANGELAH MODEL

Fig. 1 depicts the ANGELAH system model. The peculiarityof ANGELAH is the full integration between sensing andmonitoring technologies used to detect possibly dangeroussituations for the elder, along with mobile groupware collab-oration supports that enable coordination among responderswilling to engage in elder assistance. ANGELAH identifiesdifferent management roles with different responsibilities: theSensing Entity (SE) role, the Actuator Entity (AE) role,the Home Manager (HM) role, the Surveillance Center (SC)role, the Locality Manager (LM) role, and finally the LocalResponder (LR) role.

SEs represent sensors that are deployed over the elder’sliving environment. SEs continuously gather row context dataand communicate them to the HM. As depicted in Fig. 1,several SEs are deployed in elders homes, and different sensingtechnologies may be applied to monitor elder behavior andphysical conditions according to his actual needs (e.g., elder’slocation, heart-beat, environment temperature, etc).

AEs represent actuators disseminated within the elder’sapartment. AEs are controlled by the HM. They enforceappropriate actions that permit smooth monitoring of eldersconditions. For example, when the elder enters a room, RFIDtag reader (AE) turns on camera sensors and other SEs tomonitor the elder’s movement and behavior (Fig. 1).

The HM operates as a central server and is in charge ofgathering available context information from deployed SEs,aggregating this information, and detecting whether the elderis in need of help. According to the current situation, HM alsotriggers the execution of safety-related actions that typicallyinvolve the coordination among different AEs (e.g., to closepossibly open gas knobs). When an emergency situation isdetected, the HM notifies the SC of the event via an emergencynotification message, referred to as SOS throughout this paper,followed by some context information that may help a SCagent to define the emergency level and type, its causes, andthe kind of assistance the elder may be in need of (Fig. 2).

LRs represent individuals willing to provide prompt help toelders when an emergency situation occurs. LRs are staticallysubscribed to the ANGELAH service. They can be familymembers of the elder, his friends, relatives living in hisimmediate surroundings, or simply passers-by, neighborhoodcommunity representative and paid help, such as professionalcaregivers, doctors, pharmacists, etc. Each LR is characterizedby a unique User IDentifier (UID), and a profile describing hischaracteristics, such as user identity, user’s current physicallocation, user’s medical expertise, user’s history record andskills in providing assistance within the ANGELAH frame-work, and the trust SC associates with him/her. LRs’ deviceshave features that enable them to discover, join and leave eldersupport groups, to obtain the visibility of available partnersallocated in the nearby, along with their profile information,and to collaborate with them via message exchange.

5

Fig. 1. The ANGELAH model.

SC is in charge of coordinating prompt response in emer-gency situations. When SC receives a SOS message from anelder’s HM, it first obtains the visibility of all LRs allocatedin the proximity of the elder’s home. Then, on the basis ofthe visibility of LR’s profile information, SC promotes theformation of a support group composed of the best suitedLRs willing to assist the elder. Fig. 2 shows the sequence ofmessages exchanged between the SC and LRs. Upon receivinga SOS message, the SC defines the emergency type fromthe context information analysis and broadcasts a ”Call ForAssistance - CFA” message to LRs currently in the proximityof the elder. CFA messages may include information suchas personal information of the senior (e.g., age, gender, etc),the postal address of his residence, his physical and cognitivecharacteristics, the kind of assistance he is in need of, alongwith additional information (if available) describing the currentconditions of the elder (e.g., pulse). In response to the CFAmessage, LRs willing to help send back an ”AcceptanceNotification - AN” message to the SC. These reply messagescontain personal information of the volunteers (e.g., UID),their current location, and the estimated time it may take themto get to the location of the elder in need of help. SC then runsan algorithm to select the most adequate LRs. Once adequateLRs are sorted out, SC notifies them and provides them withinformation on how to access the elder’s residence and withinstructions on how to assist him/her.

It is worth noticing that composing a response group is a

rather difficult problem that call for the identification of trade-off solutions between multiple contrasting principles, such asLR’s physical location and distance with regards to the elder’shome, his medical skills, the elder pathologies and so forth,that require to solve MADM problems. Once the SC selectsa set of adequate individuals willing to provide assistance tothe elder, it notifies them and provides them with instructionson how to assist the elder. In ANGELAH, each elder supportgroup is uniquely identified by a Group IDentifier (GID) anda profile that includes information on the elder’s identity, hispathologies, his contact information and home address, and soforth.

ANGELAH is based on the locality concept both to mitigatethe complexity of the group formation problem and to reduceresponders’ intervention time. ANGELAH defines a localityas the set of all LRs located within the same network cell.ANGELAH statically associates a LM to each locality, whichis in charge of monitoring the availability of collocated LRs.It is worth noticing that several elders, possibly needingassistance may be residing in the same network locality. Inaddition, due to wireless network deployment it is also possiblefor different localities to partially overlap. LRs can freely roambetween different localities and may belong to more than onelocality at any time. Locality also plays an important rolein responders group management, because it limits the groupmanagement scope. In fact, each group can be allocated onlywithin a single locality. However, a locality can host several

6

elder support groups aiming at providing assistance to differentelders.

Fig. 2. Sequence of signaling messages exchanged between HM, SC, andLRs.

V. THE ANGELAH ARCHITECTURE

Fig. 3 portrays the ANGELAH middleware architectureimplemented on top of the Java virtual machine. As Fig. 3depicts, the layered ANGELAH architecture is mainly com-posed of three layers: the Response Management layer, theMonitoring and Assistance layer, and the Group Collaborationlayer.

Fig. 3. The ANGELAH middleware architecture.

A. Monitoring and Assistance Layer

The monitoring and assistance layer is in charge of i)integrating and managing all available sensors and actuatorsdeployed in the elder’s home; ii) gathering, aggregating,and distributing all sensed context information; iii) detectingwhether the elder needs assistance; and finally iv) controlling

Fig. 4. Monitoring and assistance layer service interaction diagram.

available actuators and alerting the surveillance center in caseof an emergency situation.

All monitoring and assistance layer support facilities are de-veloped on top of the Open Service Gateway initiative (OSGi)infrastructure. In fact, broad consensus has been reached inthe smart home research community on the important featuresof OSGi to easily integrate and manage heterogeneous homeappliances. The peculiarity of OSGi is to provide a service-oriented, component-based support that significantly simplifiessoftware lifecycle management and provides the requiredsupport to compose and frequently update sensor/actuatorinfrastructures according to the evolving elder needs.

The major interactions between the monitoring and assis-tance layer’s components are portrayed in Fig. 4. At reg-ular times, the Context Gathering Service (CGS) requestssensed data from the available sensors. The time betweenconsecutive requests hinges on sensor technical characteristics.For example, a location sensor might be frequently sam-pled to obtain a precise snapshot of elder location, whereasa temperature sensor can be sampled less frequently (e.g.,every 10 minutes) due to the large amount of time neededto change elder’s home temperature. In ANGELAH, eachsensor is statically associated to a Sensor Proxy (SP). SP isjuxtaposed between the sensors and the ANGELAH contextmanagement facilities; thus decoupling the context gatheringlogic from sensor-specific technical details and simplifyingthe context information collecting process. Once the currentsensed data are obtained from the available SEs, CGS forwardsthe obtained values to both Context Repository Service (CRS)and Context Aggregation Service (CAS). CRS persistentlymaintains context information that is required to build-upand update elder behavioral models that form the basis todetect possibly dangerous situations and to facilitate diagnosisof elder’s pathologies. On the other hand, CAS aggregatescontext information obtained from different context sourcesand detects possibly dangerous situations. Finally, ContextDistribution Service (CDS) is in charge of distributing theaggregated context data to the interested entities. For example,in case of an emergency, CDS notifies a SOS message to the

7

Emergency Response Decision Making (ERDM) instance in-stalled at SC, including information required for the formationof an elder support group, such as elder’s home address, elder’sidentity, and emergency type.

It should be emphasized here that while it is possible toconsider a self-functioning approach where upon emergencydetection, HM wirelessly broadcasts SOS messages to an ad-hoc of passers-by, with no involvement of SC similar in spiritto the outdoor emergency assistance system proposed in [20],such an approach gives rise to two major issues, namely safetyand false alerts. The first issue is highly important as weare aiming at indoor environments and the second is due tothe fact that any robust risk detection technique, which maybe installed in CAS, may still have limitations in accuratelycapturing an elder’s activities and fully understanding his be-havior, based on which anomaly can be detected. Involvementof personnel of a central surveillance center in confirming theemergency occurrence and detecting its causes and type is ofhigh importance.

B. Response Management Layer

Upon reception of an emergency notification message (e.g.,SOS message) from an elder’s HM, the response managementlayer triggers the ERDM instance to form groups composed ofindividuals willing to provide assistance to the elder. The maininteractions generated at the response management layer aredepicted in Fig. 5. As shown in the figure, ERDM coordinateswith the Proximity Service, installed at the LM of the localitywhere the elder resides, and requests the visibility of all nearbyLRs (e.g., their UID) willing to provide prompt assistance tothe elder. By coordinating with the Profile Repository (PR)at SC, ERDM can obtain further information on availablehelpers, such as their medical expertise, their history recordand skills in providing assistance within the ANGELAHframework, and the trust SC associates with them. On the basisof available information, ERDM solves a MADM problem tocompose a response group formed from the most adequateindividuals willing to help the elder. Then, ERDM coordinateswith the Group Formation Service (GFS) for promoting thenewly formed group and inviting all selected individuals tojoin it.

C. Group Collaboration Layer

The group collaboration layer evolves from the AGAPEmiddleware [20]. It provides the basic functions to compose,dissolve, and manage emergency response groups in wirelessenvironments.

When an emergency occurs, the GFS instance, installed atLM, promotes the creation of a new response group. For thispurpose, GFS coordinates with the Location-Based NamingService (LBNS) that randomly generates and assigns GIDs(and PIDs) by exploiting the Universally Unique Identifier(UUID) naming approach (Fig. 6). GFS also requests ERDMfor the elder profile information (e.g., elder’s identity, address,pathologies, emergency level, etc.) and for the list of selectedhelpers that should form the new emergency response group.Following this, GFS coordinates with the Join/Leave Manager

Fig. 5. Response management layer service interaction diagram.

Service (J/LMS) installed at helpers’ devices to invite them tojoin the group.

At regular times, the View Manager Service (VMS) creates,maintains, and disseminates views to ANGELAH group mem-bers. A view contains the list of collocated group membersalong with their profiles (context-dependent view). Each viewentry includes several data: group member PID/IP address ob-tained from PS, member’s identity and medical skills obtainedfrom PR. In addition, to cope with mobility-induced changesin group membership, VMS coordinates with PS for thenotification of arrival, departure and disconnection of a groupmember entity, and accordingly updates views. It is worthstressing out that the PS instance, installed at LM, monitors forresponders’ availability using periodic advertisement messagesfrom LRs. When the lag between consecutive advertisementmessages from a LR exceeds a threshold value, the LR isassumed to be disconnected. It is also worth noting that thethreshold value between consecutive advertisements should bechosen according to different criteria, such as the averagenumber of responders in a locality, their mobility behavior,the locality surface, and so on.

VI. INSIGHTS FOR ERDM IMPLEMENTATION

This section presents insights on the MADM algorithm usedby the ERDM instance at SC to select adequate volunteersamong a group of LRs to cope with a particular emergencysituation. Before delving into details, we list up the usednotations:• M : number of emergency levels.• =i: emergency level i (i ∈ 1, 2, · · · ,M).• θα.i: action time for emergency level =i.• τi: timeout for waiting for AN messages from LRs in

case of an emergency level =i.• γi: LR selection threshold associated with the emergency

level =i.

8

Fig. 6. Group collaboration layer service interaction diagram.

• NSi : required number of skillful LRs for emergency level

=i.• NT

i : total number of required LRs (regardless of theirskills) for emergency level =i.

• `: number of attributes associated with LRs.• Xk,j : jth attribute of LR Sk (j ∈ 1, 2, · · · , `).• Ak: vector of attributes of LR Sk (Ak =

[Xk,1, Xk,2 · · · , Xk,`]).• ωi,j : weight associated with the jth attribute in case of

emergency level =i.• Ωi: vector of weights associated with emergency level =i

(Ωi = [ωi,1, ωi,2 · · · , ωi,`]).As stated earlier, the LR selection mechanism is initiated by

ERDM based on the MADM theory. Indeed, in ANGELAH,the PR instance at SC maintains profiles of each LR; alreadysubscribed to the ANGELAH service. For each subscriberSk, a set of attributes Xk,j , j ∈ 1, 2, · · · , ` is associated.The attributes represent i) the expertise and skills of LRs, ii)their history record in providing assistance, and iii) the trustlevel SC associates with them. These attributes are constantlyupdated and maintained by SC. Table I shows a typical formatof the profile table SC maintains on registered LRs.

TABLE IA TYPICAL FORMAT OF LRS PROFILE.

LR ID 1st Attribute 2nd Attribute · · · `th AttributeSk X1 X2 X`

S1 X1.1 X1.2 · · · X1.`

S2 X2.1 X2.2 · · · X2.`

......

......

...S∝ X∝.1 X∝.2 · · · X∝.`

In ANGELAH, we assume that there are M emergencylevels defined a priori at the surveillance center. For eachemergency level =i (i ∈ 1, 2, · · · ,M) and each attribute

Xj (j ∈ 1, 2, · · · , `), SC defines a weight ωi,j as shownin Table II. Additionally, with each emergency level =i, threeparameters are associated:

• action time θα.i: minimum time within which assistanceshould be provided to the elder.

• waiting timeout τi: maximum time SC should wait for toreceive AN messages from LRs.

• acceptance threshold γi: threshold for selecting LRs.

These three parameters should be carefully set by the surveil-lance center. For example, in case of a life-threatening event(e.g., heart attack), both action time θα.i and waiting timeoutτi should be set to small values. In case of a bone fracturedue to a fall, expertise and skills become more important sothe system can set γ to high values with particular focus onthe skills-related attribute. Action time θα and timeout τ canbe set to relatively high values.

Upon receiving a SOS message, the agent in charge (at SC)first defines the corresponding emergency level based on i) theevent type (e.g., fall, faintness, heart-attack) determinable fromthe context information sent by the CDS instance (e.g., capturevideo) and ii) the profile of the senior (e.g., physical andcognitive characteristics) available at the PR instance. Let =m

denote the selected emergency level. For a timeout τm, ERDMwaits for responses from LRs. Once the system receives therequired NS

m or NTm replies, or the timeout τm expires, ERDM

sorts out the LRs based on information available in their ANmessages, such as their physical proximity and availability,using the action time Θα.m and following an analysis as willbe explained later at the end of this section. Out of the sortedLRs, those with attributes satisfying the following conditionare chosen.

Ak · ΩTm =

∑n=1

Xk,n · ωm,n ≥ γm (1)

In case the number of LRs selected at the second phase arenumerous, only an adequate number of LRs are requested toassist and are provided with information on how to access thesenior’s residence along with instructions corresponding to thedetermined emergency level =m.

In the remainder of this section, we analyze the systemresponsiveness and derive, from the analysis, conditions onthe required availability of LRs. The defined conditions areused in the selection procedure of LRs. In the analysis, thefollowing notations are used:

• θd: time elapsed since the actual occurrence of the eventtill its detection by CAS at HM. This may also includetime required by HM to confirm the anomaly either bysetting a timeout or having the senior confirm that he isindeed in need of help.

• θSoS : time required to send a SoS message along with therequired context information from CDS at HM to ERDMat SC.

• θH : time required by the agent to analyze the video and todetermine the emergency level. With this regard, agentsshould be well trained to be able to determine emergencylevels within short θH times.

9

TABLE IIEMERGENCY LEVELS AND THEIR ASSOCIATED PARAMETERS.

Emergency 1st Attribute 2nd Attribute · · · `th Attribute Action Acceptance WaitingLevel X1 X2 X` Time Threshold Timeout=1 ω1.1 ω1.2 · · · ω1.` θα.1 γ1 τ1=2 ω2.1 ω2.2 · · · ω2.` θα.2 γ2 τ2

......

......

......

......

=M ωM.1 ωM.2 · · · ωM.` θα.M γM τM

• θE : time required to exchange CFA and AN messagesbetween SC and LRs. This can be equal to the round triptime between the SC and LM.

• θERDM : time required by SC to run the ERDM algorithmfor selecting adequate LRs.

As explained earlier, a LR Sk notifies the ERDM instanceof the time he may need to reach the residence of the senior,should he be selected for the rescue task. Let ∆k denote theavailability of LR Sk in terms of time and let =m denotethe selected emergency level. Using the action time Θα.m andthe waiting timeout τm of the emergency level, LRs with timeavailability ∆k satisfying the following inequality are selected.

∆k ≤ Θα.m − τm − θd − θSoS − θH − θE − θERDM (2)

The values of θd and θSoS can be computed by ERDM fromthe context information sent by CDS.

VII. CASE STUDY: SUPPORTING ELDERS AFFECTED BYSEVERE VISION IMPAIRMENTS

To evaluate the performance of the ANGELAH frameworkin supporting in-house safety for elders, we designed and im-plemented an application tailored for elders affected by severevision impairments. Supporting vision impaired individualsis a rather challenging task, as this disability strongly limitselders’ free movements and makes individuals more subject todomestic injuries and accidental falls.

A. Deployment Settings

To secure in-house safety for elders, the underlying sensing,monitoring and positioning technologies should be customizedto the actual needs of elders. For in-house location tracking,different techniques have been recently proposed (e.g., intelli-gent floors and active badges). Whilst these technologies candetect the presence of individuals in a specific location, theyeither cannot discriminate their identities or, as a remedy tothis limitation, require users to constantly hold a device; asolution that is not efficient for seniors affected by dementia.

To cope with this limitation, we consider the use of RadioFrequency IDentification (RFID) technology for elder locationtracking. RFID tags are cost-effective and can be adhered tosenior clothes. Our envisioned network topology is as depictedin Fig. 1. The actuator and sensor infrastructure is composedof a variety of devices: video cameras, RFID readers, soundsensors, and appliances such as a smart door lock with apassword-opening function. Used cameras are standard digitalvideo cameras operating at 30 frames per second, ceiling-mounted, with vertically-oriented optical axes, fitted with

wide-angle lenses. The position and orientation of the camerasare chosen to minimize occlusion of the elder by furniture.RFID readers are also carefully placed over the house in a waythat they provide complete coverage over the entire house. It isworth stressing that, to avoid reader-to-tag interferences, twoadjacent readers should be carefully deployed in a way that thedifference between their signal strengths is less than the tag’stolerance margin [26]. By functioning as the Reader NetworkController (RNC), HM coordinates among the multiple readersusing the EPCglobal Low-Level Reader Protocol (LLRP). Toenable elder interaction with the system, microphones andspeakers are also placed in all rooms of the elder apartment.These sound I/O interfaces permit the elder to confirm to HMif he is in need of assistance upon an emergency detection bythe CAS instance in HM. Intuitively, this operation aims atavoiding false alerts to SC.

A Pentium-based PC running Gentoo Linux, J2SE 1.5,along with all ANGELAH Monitoring and Assistance layerfacilities and covering the HM role was also deployed inthe apartment. In particular, the system was appropriatelyconfigured and enabled to gather context information fromthe available sensors, i.e., RFID readers, sound sensors andcameras. Each of these entities is statically associated to aparticular SP, within the HM, in charge of sampling and pro-cessing gathered data. In particular, RFID’s SP can both detectand identify elders located in the nearby of the associatedreader, whereas camera SP can process the feedback obtainedfrom the associated camera to track the elder, to keep recordsof his motion, and to make decisions on whether an anomalyoccurred in the behavior of the senior [27]. Sound sensors areused to enhance the anomaly detection accuracy of CAS (e.g.,in case the monitored senior falls down or screams for help).It is worth noticing that the HM also installs IBM via voicein order to provide elders with a speech based interface.

While this option is not considered in our testbed, for higheraccuracy in capturing the elder’s behavior, HM needs fullvisibility of the elder’s context, including special layout of theobjects in his living environment. Such context awareness ispossible by tagging items that are frequently used by the senior(e.g., TV set, Sofa). Whenever a change occurs in the layoutof the living environment, updates of the layout is possible viaan efficient integration between the inputs from the tag readersand the cameras. The positioning of these items can be alsodone using technologies such as UWB (Ultra Wide Band) [28].However these technologies cannot provide accurate indoorpositioning and require heavy computation. Moreover, theircost is much higher compared to that of largely affordableRFID tags and readers.

10

In our testbed, we also deployed a PC running GentooLinux, J2SE 1.5, with all Response Management layer servicesand playing the SC role. In addition, by exploiting the wirelessWi-Fi infrastructure deployed in our campus in University ofBologna, we also defined several localities. In our prototype,each network cell defines a locality. In each locality, wehave also deployed a server, running Gentoo Linux, J2SE1.5, playing the LM role and supporting its PS, GFM, andVMS services. Without any loss of generality and for thesake of prototype deployment simplicity, we deployed ourprototype on top of the campus-wide Wi-Fi network. In fact,it is still difficult to conduct experimental test-beds on topof communication infrastructures managed and licensed bytelecom operators, such as GPRS, UMTS or WiMAX.

Finally, we also setup several wireless-enabled iPAQ PDAs,acting as user terminals for responders. Each device installsPersonal Java, J/LMS, and client-side components for PSand VMS. On top of ANGELAH, we also implemented anapplication prototype that provides available helpers with alertmessages when an emergency situation occurs (Fig. 11) andsupports message-oriented user collaboration and coordina-tion.

All devices OFF

Tag detected?

Wake up camera/sound sensor

Anomaly detected?

Sound I/O interface ON

Is he/she OK?

Send SOS message along with video

Analyze video Define emergency level

Timeout expires

Operations by SC Contact volunteers

Form elder support group

Operation by Tag Reader

Operations by HM

Inquiry the person if he/she is Ok.

Yes

No

Yes

No

No

Yes

Fig. 7. Major operations considered in the case study.

Fig. 7 depicts the prototype functionality. By exploitingRFID location tracking, elder’s HM obtains the visibility ofcurrent elder position in his apartment. The visibility of elder’sphysical location permits to turn on the camera and soundsensors located in the room where he is actually placed. Thisallows efficient use of energy (i.e., electricity) and permitsHM to analyze only images obtained from a single camera,

thus reducing both home network overhead and computationalload that may be induced by computer vision techniques. Itshould be noted that this operation can be also performed usingmotion sensors. However, in case the senior is sharing thehouse with other members of the family, the use of motionsensors is not efficient as they cannot distinguish the seniorfrom the other family members. In case of an emergencydetection by CAS, the AP application installed over HMturns on the speech-based I/O interface and connects it withmicrophones and speakers installed in the room where theelder is currently placed. HM first confirms with the senior ifthere is need for any kind of assistance. Awaiting a responsefrom the person, HM sets up a timeout. In case this isconfirmed by the senior or the timeout expires, HM judges theevent as an emergency situation and an alert message (SOS)is sent to SC. The audio and video captured from both thecamera and sound sensors, respectively, during the time of theemergency event are also sent to SC so that SC personnelmay look retrospectively at the moment just before the eventoccurrence to analyze the cause of the event and to decidewhether the person is in an emergency situation and what helpthe person may be in need of. LRs are contacted and an eldersupport group is formed from adequate ones. Selected LRsare provided with password to access the house of the seniorand the other volunteers are simply exempted and thanked fortheir eagerness to help. At the same time, the service providerprovides the selected volunteers with instructions on how toassist the person. Once the volunteers enter the residenceof the person in need of help, the service provider keepsmonitoring them using the cameras available at the residenceof the person and providing them with further instructionswhenever necessary.

B. Computer Vision based Emergency Detection

In practical use of the system, HM develops a normal behav-ior model of the elder, during a learning or profiling phase,at the launch of the system. This model can be developedwhen the elder is supervised by a nurse for a short period oftime (say a few days). While receiving contextual informationfrom cameras and sound sensors, HM constantly comparesthe current behavior with the normal behavior using computervision techniques. The rationale behind the choice of computervision underlies beneath the fact that computer vision offerscheap, practical and non-invasive solution in a sense that thehardware is cheap, and the user is not required to put on anydevice.

Computer vision techniques are used to track the occupantin his home environment and an AI system learns the dailylife pattern in order to detect any abnormalities such as fallsor unusual activity/inactivity patterns that can be attributed tohealth problem. In case an abnormal pattern is detected, HMsends an alert to SC. Moreover, computer vision can be usedto collect more in-depth information in order for the systemto build more accurate models of the user environment andhelp SC agents to look retrospectively to events leading to anyunusual event such as a fall. Here we describe a system wedeveloped for automatically tracking a single occupant in his

11

home environment (sitting room) and annotating his activitiesand detecting any abnormal inactivity which might be a fall[27], [29], [30].

First the occupant is tracked for an extended period of timefor HM to learn the usual entry/exit and inactivity zones in thesitting room. Once the learning is done, HM is able to detectany unusual inactivity which can indicate a matter of concernsuch as a fall and can be used as part of an alarm to alert SC.In this section, we will present the system in more details.

Fig. 8. An example of ellipse estimate (the inner ellipse) obtained by thetracking system.

Fig. 9. Smoothed trajectories and example results for learning inactivityzones.

1) Experimental Setup: The layout of the living room canbe seen in Fig. 8. The occupant might sit on the sofa to watchtelevision or on the chair to use a telephone. The sofa and thechair are referred to as normal inactivity zones to indicate thatthe occupant when sitting tends to have little global motion.The room had two entrances which also serve as exits. In [27],physical doorways used to enter the room were not entirelywithin a camera’s field of view and that is why the entries

Fig. 10. An example of fall where HM detects the occupant being inactivein an unusual location.

and exits were detected using computer vision. However inour integrated system ANGELAH, the entries and exits weredetected via RFID readers.

In the scenario studied in [27], an actor (fourth author)was instructed to perform a series of activities in the roomdesigned to emulate aspects of the way an elder might usesuch a room. He was instructed to enter the room, visit andstay for some time in one or more inactivity zones (i.e., chairor sofa), then exit the room. In an example, he was instructedto enter through the hall door, sit on the chair and use thetelephone, go and sit on the sofa and then exit through thehall door. In addition, the actor was instructed to act few fallsin different places in the room to enable evaluation of unusualinactivity. The sequences were acquired using a digital videocamera at 30 frames per second over two days of changeablelighting conditions as a result of changing weather.

2) Overhead Monitoring: As the clothing and body pos-tures were highly variable, the person’s position, and a shaperepresentation and orientation in the image plane were trackedusing an ellipse. A background subtraction technique is usedto segment moving objects (i.e., the occupant) from the staticbackground. Then a particle filtering method called IteratedLikelihood Weighting (ILW) [31] with image evidence pro-vided using background subtraction in conjunction with anadaptive background model with shadow detection. Fig. 8shows a typical estimate of the occupant whereabouts obtainedduring tracking. The tracker only lost lock in 3 sequences outof all 96 recorded sequences.

3) Context Learning and Unusual Inactivity Detection:The tracker produces temporally discretised, smoothed 2Dtrajectories in the image plane. When the speed drops below25 pixels per frame, the occupant is considered inactive.Gaussian mixture models (GMMs) were used to cluster theusual inactivity points. To obtain a Gaussian mixture withcomponents that correspond directly to inactivity zones, pe-nalized likelihood functions that encode priors on cluster scaleand shape were incorporated in the learning process. Fig. 9shows smoothed trajectories and example results for learning

12

inactivity zones. The learned model of spatial context wereused to temporally segment trajectories and to detect unusualinactivity. When the person’s speed drops below 25 pixelsper frame, he is considered inactive, and the inactivity zoneprobability density functions (pdfs) are used to check whetherthe inactivity occurred in a usual or unusual location. As thespeed at each time-step was estimated using a nite differenceover a 40-frame temporal window, an inactivity is detectedwith a delay τd = 1.6 seconds. Fig. 10 shows an exampleof fall where HM detects that the occupant is inactive in anunusual location.

C. Collaboration Management

Upon reception of a SOS message as a result of an emer-gency detection by an elder’s HM, SC promotes the formationof a new group composed of individuals currently located inphysical proximity with the elder. Responders are then invitedto join the response group.

Similar in spirit to our previous research [20], we im-plemented a prototype user interface that prevents bystanderapathy effect, which may inhibit responders from providingassistance to the elder. The designed application exploits al-ternative alerting mechanisms to render bystander interventionprompt (Fig. 11). When available LRs receive a CFA message,the application installed on their devices requires users toexplicitly acknowledge the received message. The applicationcontinuously makes the aware of the emergency notificationby emitting a sound signal with an increasing volume tillthe user accepts or declines the request. ANGELAH exploitsacknowledgments from LRs to build a list of volunteers willingto help. In addition, the application user interface providesa speech-based support built on top of IBM Via-Voice thatenables responders hand-free interaction with their devices.Fig. 11 portrays the implemented application user interfacethat provides responders with all necessary information tosecure prompt and smooth help to the elder, including pass-word to access the elder’s door-lock, ”what-to-do” and ”how-to-do” lists pertaining to tasks that need to be performed.Each task entry in the ”what-to-do” list is associated withan entry into the ”how-to-do” list that briefly instructs users,untrained or unfamiliar with the system, on how to performthe corresponding tasks.

ANGELAH introduces different forms of overhead depend-ing on the underlying management functions, such as groupcreation and context-dependent view dissemination. To evalu-ate the performance of ANGELAH, three critical quantifyingparameters are envisioned:• ANGELAH’s responsiveness in creating the elder support

groups: defined as the amount of time required forforming an elder support group. It should be emphasizedthat prompt assistance is of utmost importance in in-housesafety scenarios, particularly in case of serious patholo-gies such as heart-attacks or severe domestic injuries.

• Responders’ device battery consumption imposed by AN-GELAH’s functions: the importance of this factor stemsfrom the fact that the battery of users’ devices should beefficiently utilized not only to secure a large number of

LRs ready to assist in an emergency situation but alsoto guarantee their devices’ operability during the rescueoperations.

• Memory requirements at the user’s terminals: this factoris also critical given the fact that ANGELAH’s servicesare expected to run over portable devices, which arecharacterized by constrained resources.

We have tested ANGELAH responsiveness in scenariossimulated with NS-2 and consisting of a number of devicesrandomly deployed over the same locality, i.e., the samewireless cell (Fig. 12). To investigate the functionality ofANGELAH under different network conditions, we range thenumber of responders from 2 to 100. Our simulation settingsconsider IEEE 802.11 networks with 10 Mbps transmissionrate, random waypoint mobility pattern and speeds rangingfrom 1 to 3m/s. For the sake of simplicity, users are assumedto respond immediately to CFA messages; time needed for auser to detect alerting messages is set to zero. Admittedly, inreal-life scenarios, this time may dominate the overall groupformation time. The simulations demonstrate that ANGELAHresponsiveness is, on average, of few seconds and tend todegrade to few tens seconds when more than 60 LRs areavailable. Responsiveness degrades due to increase in packetlosses, due in turn to wireless network characteristics (e.g.,limited bandwidth and channel error rates), in addition to aminimal contribution from the computational load that comeswith the profile repository fetching the profiles of users andwith ERDM solving the MADM problem for the selection ofadequate volunteers.

The battery degradation depends mainly on the group main-tenance overhead. In particular, view dissemination is the mainfactor that contributes to battery degradation. In fact, for groupview propagation, VMS requires continuous IEEE 802.11connections which introduce typically high energy costs. Bat-tery life exhaustion augments when view dissemination rateincreases. We tried different trade-offs to ensure that theemergency tasks can be totally completed without exhaustingthe responders’ device battery. We empirically found that thetime between consecutive views dissemination should be setupfrom within the range between 10 to 15 seconds.

Fig. 13 depicts the memory requirements as seen from asingle responder device. The figure shows the overall AN-GELAH memory requirements over time. In our experiments,we tracked memory use for a responder device in a localitywhere three other responders were available. The experimentlasted for a duration of 13 minutes. All data were obtained byexploiting the JConsol profiling tool. From Fig. 13-a, the totalamount of the used heap memory varies between 0.7 MB and1.3 MB, with an average value of about 0.9 MB. In addition,non-heap memory including data, code and stack reaches astable value of about 14 MB (Fig. 13-b). From these results,it can be concluded that the ANGELAH group managementsupport can be easily installed over a PDA.

13

Fig. 11. Responders application user interface.

Fig. 12. Simulation scenario.

D. ANGELAH’s Overall Responsiveness

Admittedly, ANGELAH’s overall responsiveness2 is an im-portant metric for the system performance evaluation. How-ever, it is highly difficult to exactly determine ANGELAH’sresponsiveness. This is mainly due to the fact that the systemresponsiveness depends on unpredictable human factors, suchas the behavior of the local responders (e.g., promptness inreacting to CFA messages) and their availability in terms oftime which hinges, in turn, on their daily commitments.

In the remainder of this section, we attempt to develop amathematical analysis that can help in making an estimate

2Defined as the time elapsed since the emergency occurrence till the arrivalof the first helper to the residence of the senior in need of assistance.

(a) Heap memory.

(b) Non-heap memory.

Fig. 13. Memory use on responder devices.

of the system’s responsiveness. First, we assume that duringthe service time, LMs form a statistical profile over time onthe number of LRs that visit their locations, whether they areskillful or not, etc. When an emergency event takes place,LMs also keep history track of the responsiveness of LRs;i.e., their reaction to CFA messages via the delivery of ANmessages. From this history track, LMs can define a processfor the arrival of AN messages at SC. Additionally, from theentire database, LMs can make an estimate on the number ofvolunteers that may like to assist in an emergency situationoccurring at a particular time.

In the following, we attempt to derive the system respon-siveness using this information. As a case study, we assume

14

that the arrival process of AN messages from LRs follow aPoisson process with arrival rate λ. Let’s consider the caseof an emergency level =1, characterized by the action timeθα.1, the waiting timeout τ1, and the acceptance threshold γ1.Let also assume that to deal with the emergency level =1, SCrequires at least NS

1 volunteers with skills or NT1 ordinary

volunteers (i.e., (NS1 ≤ NT

1 ). On the assumption of Poissonprocess as a process of AN messages arrival, notificationmessages from skillful volunteers is also assumed to followPoisson process with arrival rate λs (λs ≤ λ). Based on thesame assumption, the mean number of AN messages, fromordinary and skillful volunteers respectively, arriving duringthe timeout τ1, can be computed as follows:

β = bτ1 · λc (3)βs = bτ1 · λsc (4)

The probabilities of receiving NT1 and NS

1 AN messages fromdifferent ordinary and skillful volunteers, respectively, duringthe timeout τ1 are:

PO =(τ1 · λ)NT

1 · e−λτ1

NT1 !

(5)

PS =(τ1 · λs)NS

1 · e−λsτ1

NS1 !

(6)

Again based on the Poisson process assumption of theAN message arrivals, the messages inter-arrival times aremutually independent and identically distributed. We assumethese arrivals are separated by t time units and that the firstAN message arrives t seconds after the transmission of CFAmessages. In case NT

1 AN messages are received during thetimeout τ1, the time required for receiving AN messages fromNT

1 volunteers, θO can be computed as follows:

θO = NT1 · t (7)

On the assumption of Poisson process, the probability densityfunction of t is f(t) = λe−λt. Hence, the average value of θO

is:

θavgO = NT

1 ·∫ τ1

t=0

λt · e−λtdt (8)

Considering the case when less than NT1 AN messages are

received within the timeout τ1, the average value of θO canbe expressed as:

θavgO = (1− PO)τ1 + PO ·NT

1 ·∫ τ1

t=0

λt · e−λtdt (9)

Similarly, the time required for receiving AN messages fromNS

1 skillful volunteers, θS can be computed as follows:

θavgS = (1− PS)τ1 + PS ·NS

1 ·∫ τ1

t=0

λst · e−λstdt (10)

Hence the time required for receiving AN messages sincethe transmission of CFA messages till entering the volunteerselection procedure, θAN is

θAN = max(θavgO , θavg

S ) (11)

On the other hand, out of the arriving NT1 volunteers,

only those that can meet the action time limitation θα.1 and

the acceptance threshold γ1 that will be selected. If suchvolunteers exist, the system response time, denoted as θall

can be expressed as:

θall = θd + θSoS + θH + θAN + θERDM + θE + min(∆k) (12)

Here, it should be noted that apart of min(∆k) that can beestimated from the statistical profile of LRs available at LMs,all other terms can be estimated from SC.

VIII. CONCLUSION

In this paper, we devised a middleware, dubbed ANGELAH,for supporting elders at home. Issues pertaining to elder sup-port group formation were discussed and adequate solutionswere proposed. A case study of the ANGELAH frameworkwas envisioned for elders with severe vision impairments. Inthe case study, sensors and actuators are connected to a centralunit, acting as home network manager, able of gatheringand aggregating row information from sensing sources anddetecting possibly dangerous situations based on computervision. A prototype user interface was also developed forresponders’ PDAs. The performance of ANGELAH in the casestudy was evaluated based on both computer simulations andreal-network experiments. Encouraging results were obtainedin terms of battery consumption and memory use at helpers’terminals. A mathematical analysis was also developed for thesystem’s overall responsiveness.

We are aware that the success of our prototype deeplyhinges on several aspects, ranging from socio-psychologicaland aesthetic to technical considerations. Our concern is tostimulate further research work in the area with every hopeto see robust systems, based on an efficient integration ofboth computing and telecommunication technologies, whichbetter assist elders at home and leverage their sense of safetyand autonomy. The research work, outlined in this paper, canalso define interdisciplinary projects where medical experts,caregivers, and computer/network engineers can collaborate.Indeed, as an example, caregivers can produce a library ofemergency scenarios that an elder with a particular pathologyis likely to encounter. Medical experts can define emergencylevels for each scenario, determine critical action times, andlist appropriate instructions. Engineers, in turn, can use thisinformation to define mechanisms for support group manage-ment and find adequate solutions to related MADM problems.

REFERENCES

[1] United Nations Population Division, “World Pop-ulation Prospects: the 2000 Revision”, Web Site:http://www.un.org/esa/population/publications/wpp2000/.

[2] R. Cucchiara, A. Prati, R. Vezzani, ”A Multi-Camera Vision System forFall Detection and Alarm Generation” in Expert Systems journal, vol.24, n. 5, pp. 334-345, 2007.

[3] D. Bottazzi, R. Montanari, and A. Toninelli, “Context-Aware Middle-ware for Anytime, Anywhere Social Networks,” in IEEE IntelligentSystems, Vol. 22, No. 5, Sep./Oct. 2007. pp. 23-32

[4] R.L. Keeney, H. Raiffa, “Decisions with Multiple Objectives: Pref-erences and Value Trade-offs,” Wiley, New York, 1976. Reprinted,Cambridge University Press, 1993.

[5] W. C. Mann, “The aging population and its needs,” in IEEE PervasiveComputing Mag., Vol. 3, No. 2, Apr.-Jun 2004. pp. 12-14.

15

[6] A. Soomro and D. Cavalcanti, “Opportunities and Challenges in UsingWPAN and WLAN Technologies in Medical Environments,” in IEEECommun. Mag., Feb. 2007. pp. 114-122

[7] Y. B. Choi, J. S. Krause, H. Seo, K. E. Capitan, and K. Chung,“Telemedicine in the USA: Standardization through Information Man-agement and Technical Applications,” in IEEE Commun. Mag., Vol. 44,No. 4, Apr. 2006.

[8] Y. Xiao, X. Shen, B. Sun, and L. Cai, “Security and Privacy in RFIDand Applications in Telemedicine,” in IEEE Commun. Mag., Vol. 44,No. 4, Apr. 2006.

[9] D. Hodgins, A. Bertsch, N. Post, M. Frischholz, B. Volckaerts, J.Spensley, J.M. Wasikiewicz, H. Higgins, F. von Stetten, and L. Kenney,“Healthy Aims: Developing New Medical Implants and DiagnosticEquipment,” IEEE Pervasive Computing, Vol. 7, No. 1, Jan./Mar. 2008.pp. 14 - 21.

[10] C. Y. Poon and Y. Zhang, and S. Bao, “A Novel Biometrics Methodto Secure Wireless Body Area Sensor Networks for Telemedicine andM-Health,” in IEEE Commun. Mag., Vol. 44, No. 4, Apr. 2006.

[11] S. Garawi, R. S. H. Istepanian, and M. A. Abu-Rgheff, “3G WirelessCommunications for Mobile Robotic Tele-Ultrasonography Systems,” inIEEE Commun. Mag., Vol. 44, No. 4, Apr. 2006.

[12] F. Marquis-Faulkes, S.J. McKenna, P. Gregor, and A.F. Newell,“Scenario-based drama as a tool for investigating user requirements withapplication to home monitoring for elderly people,” in Proc. 10th Int’lConf. on Human-Computer Interaction (HCI’03), Crete, Greece, 2003.

[13] A. Mihailidis, B. Carmichael, and J. Boger, “The Use of ComputerVision in an Intelligent Environment to Support Aging-in-Place, Safety,and Independence in the Home,” IEEE Trans. Information Technologyin Biomedicine, Vol. 8, No. 3, Sep. 2004. pp. 238-247

[14] M.E. Pollack, L. Brown, D. Colbry, C.E. McCarthy, C. Orosz, B.Peintner, S. Ramakrishnane, and I. Tsamardinos, “Autominder: An in-telligent cognitive orthotic system for people with memory impairment,”in Robotics and Autonomous Systems, Vol. 44, No. 3-4, Sep. 2003. pp273-282.

[15] B.P. Lo, J.L. Wang, and G.Z. Yang, “From Imaging Networks toBehavior Profiling: Ubiquitous Sensing for Managed Homecare of The.Elderly,” in Proc. 3rd Int’l Conf. on Pervasive Computing, Munich,Germany, May 2005.

[16] A.M. Tabar, A. Keshavarz, and H. Aghajan, “Smart Home Care Networkusing Sensor Fusion and Distributed Vision-based Reasoning,” in Proc.4th ACM Int’l Workshop on Video Surveillance and Sensor Network,Santa Barbara, USA, Nov. 2006.

[17] K. Z. Haigh, L. M. Kiff, J. Myers, V. Guralnik, C. W. Geib, J. Phelps, andT. Wagner, “The Independent Lifestyle Assistant (I.L.S.A.): AI LessonsLearned,” in Proc. 16th Innovative Applications of Artificial IntelligenceConf. (IAAI’04), San Jose, CA., Jul. 2004.

[18] F.D. Mola, G. Cabri, N. Muratori, R. Quitadamo, and F. Zambonelli,“The UbiMedic Framework to Support Medical Emergencies by Ubiq-uitous Computing,” in Int’l Trans. Systems Science & Applications, Vol.1, No. 1, 2006. pp. 15-26

[19] C. Pesto, L. Selavo, Y. Chen, J. Ko, J.H. Lim, A. Terzis, A. Watt, J.Jeng, B. Chen, K. Lorincz, and M. Welsh, “Wireless Medical SensorNetworks in Emergency Response: Implementation and Pilot Results,”In Proc. 2008 IEEE Int’l Conf. on Technologies for Homeland Security,Waltham, MA, May 2008.

[20] D. Bottazzi, A. Corradi, and R. Montanari, “Context-Aware MiddlewareSolutions for Anytime and Anywhere Emergency Assistance to ElderlyPeople,” in IEEE Commun. Mag., Vol. 44, No. 4, Apr. 2006. pp. 82-90

[21] A. Mihailidis, G. R. Fernie, and J. C. Barbenel, “The Use of ArtificialIntelligence in the Design of an Intelligent Cognitive Orthosis for Peoplewith Dementia,” in Assistive Technology, Vol. 13, No. 1, 2001. pp. 2339.

[22] M. Morris and J. Lundell, , “Ubiquitous Computing for CognitiveDecline: Findings From Intels Proactive Health Research,” WA: IntelCorp., Seattle, 2003.

[23] P. Bellavista, D. Bottazzi, A. Corradi, and R. Montanari, “Challenges,opportunities and solutions for ubiquitous eldercare,” DEIS Universityof Bologna Technical Report, Oct. 2005.

[24] A. Mihailidis and G. Fernie, “Context-aware Assistive Devices for OlderAdults with Dementia,” in Gerontechnology, Vol. 2, No. 2, 2002. pp.173-189.

[25] A. K. Dey, “Understanding and Using Context,” in Personal and Ubiq-uitous Computing Journal, Vol. 5, No. 1, 2001. pp. 4-7.

[26] P. Krishna and D. Husak, “RFID Infrastructure,” in IEEE Applications& Practice, Vol. 45, No. 9, Sep. 2007. pp. 4-10

[27] S. J. McKenna and H. Nait-Charif, “Summarising Contextual Activityand Detecting Unusual Inactivity in a Supportive Home Environment,”

in Pattern Analysis and Applications, Vol. 7, No. 4, Dec. 2004, pp.386-401.

[28] K. Pahlavan, F. Akgul, M. Heidari, A. Hatami, J. Elwell, and R. Tingley,“Indoor Geolocation in the Absence of Direct Path,” in IEEE WirelessCommun. Mag., Vol. 13, No. 6, Dec. 2006. pp. 50-58

[29] H. Nait-Charif and S. J. McKenna, “Activity Summarisation and FallDetection in a Supportive Home Environment,” in Proc. Int’l Conf. onPattern Recognition (ICPR), Cambridge, Aug. 2004.

[30] S. J. McKenna and H. Nait-Charif, “Learning Spatial Context fromTracking using Penalized Likelihood Estimation,” in Proc. Int’l Conf.on Pattern Recognition (ICPR), Cambridge, Aug. 2004.

[31] H. Nait-Charif and S. J. McKenna, “Head Tracking and Action Recog-nition in a Smart Meeting Room,” in Proc. 4th IEEE Int’l Workshopon Performance Evaluation of Tracking and Surveillance, Int’l Conf. onComputer Vision Systems (PETS-ICVS), Graz, Austria, pp. 24-31, Mar.2003.